Add blob provider benchmark for auxiliary key fallback

[sffc]

To evaluate solutions for #2699

[sffc]

I added some #[inline(never)]s and ran pprof. Here's the weighted call tree for provider/auxkey/fallback/sr-ME-x-ym0d/v2

Reproduction instructions:

1. Checkout https://github.com/sffc/omnicu/tree/archive/auxkey-bench-pprof
2. From provider/blob, run cargo bench --all-features auxkey/fallback -- --profile-time 10
3. Install pprof if you haven't already done so: go install github.com/google/pprof@latest
4. Then open pprof like this: ~/go/bin/pprof -http=$hostname:12000 ../../target/release/deps/auxkey_bench-105bf790d34625b3 ../../target/criterion/provider_auxkey_fallback_sr-ME-x-ym0d_v2/profile/profile.pb

[sffc]

Summary of the main contributors to data loading performance, according to the pprof visualization:

• 31% zerotrie reader function
• 24% writeable impl walking the DataLocale struct
• 20% fallback algorithm (step_language and run_fallback, which may include some data access)
• 8% deserialization
• 7% blob data access / indexing
• 4% fallback data access (likely subtags ZeroMap2d)
• 3% memcmp

Both of the top two contributors are things I've already spent quite a bit of time optimizing, which is good because they clearly have the most impact. However, there's always room for improvement.